CoRL 2025 论文列表

We present PercepSkin, a pre-trained encoder for magnetic skin sensors distributed across the fingertips, phalanges, and palm of a dexterous robot hand. Magnetic tactile skins offer a flexible form factor for hand-wide coverage with fast response times, in contrast to vision-based tactile sensors that are restricted to the fingertips and limited by bandwidth. Full hand tactile perception is crucial for robot dexterity. However, a lack of general-purpose models, challenges with interpreting magnetic flux and calibration have limited the adoption of these sensors. PercepSkin, given a history of kinematic and tactile sensing across a hand, outputs a latent tactile embedding that can be used in any downstream task. The encoder is self-supervised via self-distillation on a variety of unlabeled hand-object interactions using an Allegro hand sensorized with Xela uSkin. In experiments across several benchmark tasks, from state estimation to policy learning, we find that pretrained PercepSkin representations are both sample efficient in learning downstream tasks and improve task performance by over 41% compared to prior work and over 56% compared to end-to-end learning.

Dynamic rotational maneuvers, such as front flips, inherently involve large angular momentum generation and intense impact forces, presenting major challenges for reinforcement learning and sim-to-real transfer. In this work, we propose a general framework for learning and deploying impact-rich, rotation-intensive behaviors through centroidal velocity-based rewards and actuator-aware sim-to-real techniques. We identify that conventional link-level reward formulations fail to induce true whole-body rotation and introduce a centroidal angular velocity reward that accurately captures system-wide rotational dynamics. To bridge the sim-to-real gap under extreme conditions, we model motor operating regions (MOR) and apply transmission load regularization to ensure realistic torque commands and mechanical robustness. Using the one-leg hopper front flip as a representative case study, we demonstrate the first successful hardware realization of a full front flip. Our results highlight that incorporating centroidal dynamics and actuator constraints is critical for reliably executing highly dynamic motions.

Recently, 2D vision-language-action (VLA) models have made significant strides in multi-task manipulation. However, these models struggle to reason about 3D spatial relationships from 2D image inputs. Although an increasing number of 3D approaches explicitly integrate 3D information, they encounter challenges such as limited availability of large-scale 3D datasets and loss of spatial information during input processing. Meanwhile, existing policies typically focus on the perception-to-action learning paradigm, lacking an explicit understanding of the spatial and temporal relationships between the robot and its environment. To address this, we propose 3DS-VLA, which enhances pretrained 2D vision-language models (VLMs) with comprehensive 3D awareness, enabling the prediction of robust end-effector poses. Specifically, we enable a 2D vision encoder to encode both 2D images and 3D spatial observation by introducing a 2D-to-3D positional alignment mechanism. This allows 3DS-VLA to leverage the large-scale pre-trained knowledge of the VLM for effective reasoning in complex 3D robotic environments. Furthermore, to better understand the spatiotemporal relationship between 3D observations and robot behavior, we guide the model to learn the introduced sequential 3D spatial constraints, which define affordance-relevant 3D keypoints on objects, ensuring robust interactions. Experiments in simulated and real-world demonstrate that 3DS-VLA outperforms previous state-of-the-art policies and showcase its generalizable capabilities across multi-task, multi-embodiment, and diverse environmental settings.

Clear communication of robot intent fosters transparency and interpretability in physical human-robot interaction (pHRI), particularly during assistive tasks involving direct human-robot contact. We introduce CoRI, a pipeline that automatically generates natural language communication of a robot's upcoming actions directly from its motion plan and visual perception. Our pipeline first processes the robot's image view to identify human poses and key environmental features. It then encodes the planned 3D spatial trajectory (including velocity and force) onto this view, visually grounding the path and its dynamics. CoRI queries a vision-language model with this visual representation to interpret the planned action within the visual context before generating concise, user-directed statements, without relying on task-specific information. Results from a user study involving robot-assisted feeding, bathing, and shaving tasks across two different robots indicate that CoRI leads to statistically significant difference in communication clarity compared to a baseline communication strategy. Specifically, CoRI effectively conveys not only the robot's high-level intentions but also crucial details about its motion and any collaborative user action needed.

Comprehensive, unbiased, and comparable evaluation of modern generalist policies is uniquely challenging: existing approaches for robot benchmarking typically rely on heavy standardization, either by specifying fixed evaluation tasks and environments, or by hosting centralized "robot challenges", and do not readily scale to evaluating generalist policies across a broad range of tasks and environments. In this work, we propose RoboArena, a new approach for scalable evaluation of generalist robot policies in the real world. Instead of standardizing evaluations around fixed tasks, environments, or locations, we propose to crowd-source evaluations across a distributed network of evaluators. Importantly, evaluators can freely choose the tasks and environments they evaluate on, enabling easy scaling of diversity, but they are required to perform double-blind evaluations over pairs of policies. Then, by aggregating preference feedback from pairwise comparisons across diverse tasks and environments, we can derive a ranking of policies. We instantiate our approach across a network of evaluators at seven academic institutions using the DROID robot platform. Through more than 600 pairwise real-robot evaluation episodes across seven generalist policies, we demonstrate that our crowd-sourced approach can more accurately rank the performance of existing generalist policies than conventional, centralized evaluation approaches, while being more scalable, resilient, and trustworthy. We open our evaluation network to the community and hope that it can enable more accessible comparisons of generalist robot policies.

Image-based behaviour cloning leverages demonstrations captured from ubiquitous RGB cameras, enabling impressive visuomotor performance. However, it remains constrained by the cost of collecting sufficiently diverse demonstrations, especially for generalizing across workspace variations. We propose MirrorDuo, a mirroring-based formulation that operates on image, proprioception, and full 6-DoF end-effector action tuples, generating a mirrored counterpart for each original demonstration, effectively achieving ``collect one, get one for free." It can be applied as a data augmentation strategy for existing learning pipelines, such as standard behaviour cloning or diffusion policy, or as a structural prior for reflection-equivariant policy networks. By leveraging the overlap between the original and mirrored domains, MirrorDuo achieves significantly improved performance under the same data budget when demonstrations are evenly distributed across both sides of the workspace. When demonstrations are confined to one side, MirrorDuo enables efficient skill transfer to the mirrored workspace with as few as zero or just 5 demonstrations in the target arrangement.

Robot chain-of-thought reasoning (CoT) -- wherein a model predicts helpful intermediate representations before choosing actions -- provides an effective method for improving the generalization and performance of robot policies, especially vision-language-action models (VLAs). While such approaches have been shown to improve performance and generalization, they suffer from core limitations, like needing specialized robot reasoning data and slow inference speeds. To design new robot reasoning approaches that address these issues, a more complete characterization of why reasoning helps policy performance is critical. We hypothesize several mechanisms by which robot reasoning improves policies -- (1) better representation learning, (2) improved learning curricularization, and (3) increased expressivity -- then devise simple variants of robot CoT reasoning to isolate and test each one. We find that learning to generate reasonings does lead to better VLA representations, while attending to the reasonings aids in actually leveraging these features for improved action prediction. Our results provide us with a better understanding of why CoT reasoning helps VLAs, which we use to introduce two simple and lightweight alternative recipes for robot reasoning. Our proposed approaches achieve significant performance gains over non-reasoning policies, state-of-the-art results on the LIBERO-90 benchmark, and a 3x inference speedup compared to standard robot reasoning.

In this paper, we propose AimBot, a lightweight visual augmentation technique that provides explicit spatial cues to improve visuomotor policy learning in robotic manipulation. AimBot overlays shooting lines and scope reticles onto multi-view RGB images, offering auxiliary visual guidance that encodes the end-effector's state. The overlays are computed from depth images, camera extrinsics, and the current end-effector pose, explicitly conveying spatial relationships between the gripper and objects in the scene. AimBot incurs minimal computational overhead (less than 1 ms) and requires no changes to model architectures, as it simply replaces original RGB images with augmented counterparts. Despite its simplicity, our results show that AimBot consistently improves the performance of various visuomotor policies in both simulation and real-world settings, highlighting the benefits of spatially grounded visual feedback. More videos can be found at https://aimbot-reticle.github.io/

This paper presents Latent Adaptive Planner (LAP), a novel approach for dynamic nonprehensile manipulation tasks that formulates planning as latent space inference, effectively learned from human demonstration videos. Our method addresses key challenges in visuomotor policy learning through a principled variational replanning framework that maintains temporal consistency while efficiently adapting to environmental changes. LAP employs Bayesian updating in latent space to incrementally refine plans as new observations become available, striking an optimal balance between computational efficiency and real-time adaptability. We bridge the embodiment gap between humans and robots through model-based proportional mapping that regenerates accurate kinematic-dynamic joint states and object positions from human demonstrations. Experimental evaluations across multiple complex manipulation benchmarks demonstrate that LAP achieves state-of-the-art performance, outperforming existing approaches in success rate, trajectory smoothness, and energy efficiency, particularly in dynamic adaptation scenarios. Our approach enables robots to perform complex interactions with human-like adaptability while providing an expandable framework applicable to diverse robotic platforms using the same human demonstration videos.

Sampling-based motion planning methods, while effective in high-dimensional spaces, often suffer from inefficiencies due to irregular sampling distributions, leading to suboptimal exploration of the configuration space. In this paper, we propose an approach that enhances the efficiency of these methods by utilizing low-discrepancy distributions generated through Message-Passing Monte Carlo (MPMC). MPMC leverages Graph Neural Networks (GNNs) to generate point sets that uniformly cover the space, with uniformity assessed using the the $\mathcal{L}_p$-discrepancy measure, which quantifies the irregularity of sample distributions. By improving the uniformity of the point sets, our approach significantly reduces computational overhead and the number of samples required for solving motion planning problems. Experimental results demonstrate that our method outperforms traditional sampling techniques in terms of planning efficiency.

Assistive teleoperation, where control is shared between a human and a robot, enables efficient and intuitive human-robot collaboration in diverse and unstructured environments. A central challenge in real-world assistive teleoperation is for the robot to infer a wide range of human intentions from user control inputs and to assist users with correct actions. Existing methods are either confined to simple, predefined scenarios or restricted to task-specific data distributions at training, limiting their support for real-world assistance. We introduce Casper, an assistive teleoperation system that leverages commonsense knowledge embedded in pre-trained visual language models (VLMs) for real-time intent inference and flexible skill execution. Casper incorporates an open-world perception module for a generalized understanding of novel objects and scenes, a VLM-powered intent inference mechanism that leverages commonsense reasoning to interpret snippets of teleoperated user input, and a skill library that expands the scope of prior assistive teleoperation systems to support diverse, long-horizon mobile manipulation tasks. Extensive empirical evaluation, including human studies and system ablations, demonstrates that Casper improves task performance, reduces human cognitive load, and achieves higher user satisfaction than direct teleoperation and assistive teleoperation baselines. More information is available at https://casper-corl25.github.io/

Humans excel at bimanual assembly tasks by adapting to rich tactile feedback—a capability that remains difficult to replicate in robots through behavioral cloning alone, due to the suboptimality and limited diversity of human demonstrations. In this work, we present VT-Refine, a visuo-tactile policy learning framework that combines real-world demonstrations, high-fidelity tactile simulation, and reinforcement learning to tackle precise, contact-rich bimanual assembly. We begin by training a diffusion policy on a small set of demonstrations using synchronized visual and tactile inputs. This policy is then transferred to a simulated digital twin equipped with simulated tactile sensors and further refined via large-scale reinforcement learning to enhance robustness and generalization. To enable accurate sim-to-real transfer, we leverage high-resolution piezoresistive tactile sensors that provide normal force signals and can be realistically modeled in parallel using GPU-accelerated simulation. Experimental results show that VT-Refine improves assembly performance in both simulation and the real world by increasing data diversity and enabling more effective policy fine-tuning. Our project page is available at https://vt-refine.github.io/ .

Imitation learning enables intelligent systems to acquire complex behaviors with minimal supervision. However, existing methods often focus on short-horizon skills, require large datasets, and struggle to solve long-horizon tasks or generalize across task variations and distribution shifts. We propose a novel neuro-symbolic framework that jointly learns continuous control policies and symbolic domain abstractions from a few skill demonstrations. Our method abstracts high-level task structures into a graph, discovers symbolic rules via an Answer Set Programming solver, and trains low-level controllers using diffusion policy imitation learning. A high-level oracle filters task-relevant information to focus each controller on a minimal observation and action space. Our graph-based neuro-symbolic framework enables capturing complex state transitions, including non-spatial and temporal relations, that data-driven learning or clustering techniques often fail to discover in limited demonstration datasets. We validate our approach in six domains that involve four robotic arms, Stacking, Kitchen, Assembly, and Towers of Hanoi environments, and a distinct Automated Forklift domain with two environments. The results demonstrate high data efficiency with as few as five skill demonstrations, strong zero- and few-shot generalizations, and interpretable decision making. Our code is publicly available.

Autonomous skill discovery aims to enable robots to acquire diverse be-haviors without explicit supervision. Learning such behaviors directly on physical hardware remains challenging due to safety and data efficiency constraints. Existing methods, including Quality-Diversity Actor-Critic (QDAC), require manually defined skill spaces and carefully tuned heuristics, limiting real-world applicability. We propose Unsupervised Real-world Skill Acquisition (URSA), an extension of QDAC that enables robots to autonomously discover and master diverse, high-performing skills directly in the real world. We demonstrate that URSA successfully discovers diverse locomotion skills on a Unitree A1 quadruped in both simulation and the real world. Our approach supports both heuristic-driven skill discovery and fully unsupervised settings. We also show that the learn skill repertoire can be reused for downstream tasks such as real-world damage adaptation, where URSA outperforms all baselines in 5 out of 9 simulated and 3 out of 5 real-world damage scenarios. Our results establish a new framework for real-world robot learning that enables continuous skill discovery with limited human intervention, representing a significant step toward more autonomous and adaptable robotic systems.

Tactile-aware robot learning faces critical challenges in data collection and representation due to data scarcity and sparsity, and the absence of force feedback in existing systems. To address these limitations, we introduce a tactile robot learning system with both hardware and algorithm innovations. We present exUMI, an extensible data collection device that enhances the vanilla UMI with robust proprioception (via AR MoCap and rotary encoder), modular visuo-tactile sensing, and automated calibration, achieving 100% data usability. Building on an efficient collection of over 1 M tactile frames, we propose Tactile Prediction Pretraining (TPP), a representation learning framework through action-aware temporal tactile prediction, capturing contact dynamics and mitigates tactile sparsity. Real-world experiments show that TPP outperforms traditional tactile imitation learning. Our work bridges the gap between human tactile intuition and robot learning through co-designed hardware and algorithms, offering open-source resources to advance contact-rich manipulation research.

This paper introduces IRIS, an Immersive Robot Interaction System leveraging Extended Reality (XR). Existing XR-based systems enable efficient data collection but are often challenging to reproduce and reuse due to their specificity to particular robots, objects, simulators, and environments. IRIS addresses these issues by supporting immersive interaction and data collection across diverse simulators and real-world scenarios. It visualizes arbitrary rigid and deformable objects, robots from simulation, and integrates real-time sensor-generated point clouds for real-world applications. Additionally, IRIS enhances collaborative capabilities by enabling multiple users to simultaneously interact within the same virtual scene. Extensive experiments demonstrate that IRIS offers efficient and intuitive data collection in both simulated and real-world settings.

We present Latent Theory of Mind (LatentToM), a decentralized diffusion policy architecture for collaborative robot manipulation. Our policy allows multiple manipulators with their own perception and computation to collaborate with each other towards a common task goal with or without explicit communication. Our key innovation lies in allowing each agent to maintain two latent representations: an ego embedding specific to the robot, and a consensus embedding trained to be common to both robots, despite their different sensor streams and poses. We further let each robot train a decoder to infer the other robot's ego embedding from their consensus embedding, akin to "theory of mind" in latent space. Training occurs centrally, with all the policies' consensus encoders supervised by a loss inspired by sheaf theory, a mathematical theory for clustering data on a topological manifold. Specifically, we introduce a first-order cohomology loss to enforce sheaf-consistent alignment of the consensus embeddings. To preserve the expressiveness of the consensus embedding, we further propose structural constraints based on theory of mind and a directional consensus mechanism. Execution can be fully distributed, requiring no explicit communication between policies. In which case, the information is exchanged implicitly through each robot's sensor stream by observing the actions of the other robots and their effects on the scene. Alternatively, execution can leverage direct communication to share the robots' consensus embeddings, where the embeddings are shared once during each inference step and are aligned using the sheaf Laplacian. While we tested our method using two manipulators, our approach can naturally be extended to an arbitrary number of agents. In our hardware experiments, LatentToM outperforms a naive decentralized diffusion baseline, and shows comparable performance with a state-of-the-art centralized diffusion policy for bi-manual manipulation. Additionally, we show that LatentToM is naturally robust to temporary robot failure or delays, while a centralized policy may fail.

Dexterous robotic hands enable versatile interactions through the flexibility and adaptability of a multi-finger setup, allowing for a wise range of task-specific grasp configurations in diverse environments. However, access to diverse and high-quality grasp data is essential to fully exploit the capabilities of dexterous hands, be it to train grasp prediction models from point clouds, train manipulation policies, or to support high-level task planning with a broader range of action options. Existing approaches for dataset generation rely on sampling-based algorithms or simplified force-closure analysis, which tend to converge to power grasps and often exhibit limited diversity. In this work, we propose a method to synthesize large-scale, diverse, and physically feasible grasps that additionally go beyond simple power grasps to more refined manipulation, such as pinches or tri-finger precision grasps. We introduce a rigorous differentiable energy formulation of force closure, implicitly defined through a Quadratic Program (QP). In addition, we present an adjusted optimization method (MALA*) that improves performance by dynamically rejecting gradient steps based on the global sample distribution. We extensively evaluate our approach and demonstrate significant improvements in both grasp diversity and the stability of final grasp predictions. Finally, we provide a new, large-scale grasp dataset for the 5'700 objects from DexGraspNet, consisting of five different grippers and three different grasp types.

Multi-part assembly poses significant challenges for robotic systems to execute long-horizon, contact-rich manipulation with generalization across complex geometries. We present a dual-arm robotic system capable of end-to-end planning and control for autonomous assembly of general multi-part objects. For planning over long horizons, we develop hierarchies of precedence, sequence, grasp, and motion planning with automated fixture generation, enabling general multi-step assembly on any dual-arm robots. The planner is made efficient through a parallelizable design and is optimized for downstream control stability. For contact-rich assembly steps, we propose a lightweight reinforcement learning framework that trains generalist policies across object geometries, assembly directions, and grasp poses, guided by equivaraiance and residual actions obtained from the plan. These policies transfer zero-shot to the real world and achieve 80% success rates. For systematic evaluation, we propose a benchmark suite of multi-part assemblies resembling industrial and daily objects across diverse categories and geometries. By integrating efficient global planning and robust local control, we demonstrate the first system to achieve complete and generalizable real-world multi-part assembly without domain knowledge or human demonstrations.